AI didn't collapse the cost of coding. It collapsed the cost of finding out.

The claim, and the messier evidence behind it

The line making the rounds in founder and venture circles is blunt: AI has collapsed the cost of building software, so the only real advantage left is finding customers before anyone else does. Parts of that claim hold up well. The part that doesn't is the assumption everyone seems to make next: that AI has made engineers themselves dramatically faster at their jobs. The research on that specific question is far less settled than the anecdotes suggest.

METR, a research nonprofit that studies AI capabilities, ran a randomised controlled trial between February and June 2025 with experienced open-source developers working on their own real repositories. Tasks were randomly assigned to an AI-allowed or AI-disallowed condition, and developers logged how long each one took. The result: using AI tools made these developers about 19% slower on average, not faster, even though the same developers believed AI had sped them up by roughly 20%. METR then ran a follow-up study, starting in August 2025, specifically to see whether that had changed as tools improved. It has since published an unusually candid account of why the follow-up data can’t be trusted at face value: a growing share of developers refused to take part in tasks without AI at all, and among those who did participate, 30 to 50% said they were avoiding submitting tasks they didn’t want to attempt "the old way." Both effects bias the study toward measuring exactly the developers and tasks where AI helps least. METR’s own conclusion is that developers are probably faster now than the original study found, but that its newer data is, in its words, "only very weak evidence for the size of this increase."

Stack Overflow's 2025 Developer Survey, fielded across more than 49,000 developers, tells a related story from a different angle. 84% of respondents said they use or plan to use AI tools in their work, up from 76% the year before. But trust in the accuracy of what those tools produce fell to 29%, down from 40% in 2024, and 46% said they actively distrust the output. Adoption is close to universal. Confidence in the result is falling, not rising, even as usage climbs.

What actually got cheap

None of this means nothing has changed. It means the change isn't mainly happening at the level of an individual engineer typing faster. It's happening a step earlier: in how cheaply someone can find out whether an idea is worth building at all.

A useful illustration that's genuinely fintech, and squarely UK/EU relevant, is Kashimi, a Vilnius-based open banking payment infrastructure provider. Incorporated in November 2024, it had, within about eleven months, built a specialist team of roughly ten open banking engineers, signed its first client integrations, and raised $1.36 million in pre-seed funding backed in part by Plug and Play. Kashimi's own pitch to its clients, banks, e-money institutions and other regulated payment providers, is that its infrastructure lets them launch new account-to-account payment products in weeks rather than months. It's worth being precise about what that does and doesn't prove: Kashimi's own account emphasises its open banking infrastructure and engineering execution, not AI coding tools specifically, so it isn't evidence that AI caused the speed. What it does show is that the broader compression this article is describing is visible inside regulated fintech infrastructure itself, not only in general-purpose app builders.

The same compression shows up at a market level too, and here the relevance to UK and Western European founders is more direct: a large share of the companies in this dataset are exactly the kind of small technology and fintech businesses Aulay's clients encounter as competitors, acquisition targets or portfolio companies. Stripe's own data on companies formed through its Atlas incorporation service found that more of the businesses started in 2025 reached $100,000 in revenue within their first six months than in 2024 (56% more), and got there faster: a median of 108 days rather than 121. The average number of paying customers a new company had acquired within six months rose by more than half. Stripe attributes this partly to AI-assisted development and partly to improvements in payments and compliance infrastructure that founders no longer have to build themselves, not to AI coding tools alone, which matters: the collapse in cost is as much about better financial and regulatory plumbing, the same kind of infrastructure Kashimi itself provides, as it is about AI writing code.

The same data carries a genuine caveat, and it's worth stating plainly rather than skipping past it. Solo-founded companies now account for 63% of new C corporations formed through Stripe Atlas, an all-time high. But the gap between typical outcomes and the best ones is widening, not narrowing: median first-half revenue for solo-founded startups fell 23% year over year in 2025, while revenue at the top decile rose 19%. The honest reading is that cheaper tools help most when a founder already has an unusually strong idea, some existing distribution or genuine domain expertise, and help considerably less otherwise. AI lowering the cost of building doesn't lower the cost of having a good idea.

The framework this actually supports

Put together, the honest version of the claim is narrower than the slogan suggests. The loop that used to run for months, and cost real money before anyone found out whether an idea justified the first hire, now compresses to weeks or days at its earliest stages: idea, prototype, a real market test, iterate. What hasn’t compressed is everything downstream of "does anyone want this," once the answer concerns a regulated financial product rather than a to-do-list app. Getting a prototype in front of a handful of test users is close to free now. Getting it in front of the users who actually matter, in a form a bank, an insurer or a supervisor will engage with, still runs into the same constraints that apply regardless of how quickly the software was assembled: licensing, data protection, securing a first paying institutional client, and the operational-resilience obligations that regimes such as DORA impose on regulated counterparties no matter how the underlying system was built.

Why this cuts differently for a fintech than for a consumer app

For a product bought by an individual professional or a small team, self-serve signup, a credit card, no procurement process, the bottleneck really was mostly "can we build this fast enough," and AI genuinely dissolves a good part of that constraint. For a fintech aimed at banks, insurers, asset managers or private-equity-backed portfolios, engineering speed was rarely the true constraint in the first place. It was trust, compliance and a sales cycle measured in quarters rather than weeks. What’s changed is that the cheap, disposable early prototype now exists before any of that slower validation work even begins.

That has a practical implication for where a founder, or a bank running an internal venture, should spend scarce capital and credibility. It shouldn't go into polishing a demo, which is now nearly free to produce and iterate on. It should go into the harder and unavoidably slower work: finding the first counterparty willing to say yes, and proving the idea survives contact with a real institution's risk, compliance and procurement processes. Teams that treat the prototype stage as the expensive part, out of habit built up over the last decade, are misallocating the one resource that genuinely didn't get cheaper.

The bottleneck moved, but the slogan oversimplifies where

"Distribution is the new moat" is the phrase most often attached to this shift, and it’s directionally right while being analytically thin. Lou Shipley, a senior lecturer at Harvard Business School, has made the point directly: the popular version of this argument treats distribution as though it were something a company owns outright, when for most companies it isn’t. It’s borrowed: a startup selling through a bank’s branch network is reaching the bank’s customers, not its own; a company selling through an app store or platform is reaching that platform’s audience; a firm relying on a channel partner is reaching the partner’s contacts. Whoever actually controls that access can change the terms, charge more for it, or hand it to a competitor. That makes it valuable, but it isn’t a moat in the traditional sense: it’s leverage that still has to be earned or paid for continuously, in much the same way good engineers used to be the scarce resource worth fighting over. A future piece in this series scores exactly this trade-off across five dimensions together, technology, distribution, trust, economics and execution, because no single one of them decides the outcome on its own.

Aulay's advice to a client running an internal fintech build, or a founder weighing one, is consistent with the evidence above rather than the slogan: treat the early prototype as disposable, because it genuinely now is, and treat the market test as the real expense of time and credibility it has always been. Compress the diagnose-and-prototype phase deliberately, and redirect the time that frees up into the slower, unavoidable work: finding one real institutional counterparty willing to commit, and proving the idea holds up under the regulatory and operational scrutiny that a cheap demo was never going to face in the first place.

If you're weighing whether an internal fintech initiative, or a venture you're building outside one, is actually ready to test against a real customer rather than another internal demo, get in touch and we'll talk through where the build currently stands.